ReturnFromMoon: change periapsis constraint

[lamont-granquist]

removing normalization makes it converge much faster.

slight impact on accuracy but cuts the time by over 50%.

also add an always-on-optguard test so i make sure not to screw up the jacobians.